开源模型选型指南

327 个开源模型，覆盖主流厂商。按参数规模、架构类型、VRAM 需求筛选，找到最适合你硬件条件的开源模型。

Large(>150B)

Medium(40-150B)

Small(4-40B)

135

Tiny(<=4B)

智能指数 vs VRAM 需求 (Q4)

气泡大小 = 参数量，右下 = 高智能低显存 (部署性价比)，近 18 个月模型，共 265 个数据点 · 更新于 2026/06/27

TrakToken

https://www.traktoken.com

大小:

架构:

硬件:

共 327 个开源模型

排序:

VRAM 兼容性检查器

选择你的显卡和量化级别，查看可运行的开源模型列表

NVIDIA RTX 4090x 1 =24GB总显存(Q4 量化)

可运行 166 个开源模型

Qwen3.6 27B (Reasoning)阿里云

IQ 37.117GB

Qwen3.5 27B (Reasoning)阿里云

IQ 33.817GB

Qwen3.6 35B A3B (Reasoning)阿里云MoE

IQ 31.622GB

Gemma 4 31B (Reasoning)Google

IQ 29.418GB

Qwen3.5 27B (Non-reasoning)阿里云

IQ 29.317GB

Qwen3.5 35B A3B (Reasoning)阿里云MoE

IQ 29.322GB

Qwen3.6 27B (Non-reasoning)阿里云

IQ 29.317GB

Gemma 4 26B A4B (Reasoning)GoogleMoE

IQ 25.715GB

Qwen3.5 9B (Reasoning)阿里云

IQ 25.06GB

Gemma 4 31B (Non-reasoning)Google

IQ 24.818GB

Qwen3.6 35B A3B (Non-reasoning)阿里云MoE

IQ 24.222GB

Qwen3.5 35B A3B (Non-reasoning)阿里云MoE

IQ 23.422GB

EXAONE 4.5 33BLG AI Research

IQ 23.021GB

GLM-4.7-Flash (Reasoning)Z AI (智谱 AI)MoE

IQ 22.919GB

Gemma 4 12B (Reasoning)Google

IQ 22.07GB

Nemotron Cascade 2 30B A3BNVIDIAMoE

IQ 21.319GB

Apriel-v1.5-15B-ThinkerServiceNow

IQ 21.29GB

Apriel-v1.6-15B-ThinkerServiceNow

IQ 20.59GB

Qwen3.5 9B (Non-reasoning)阿里云

IQ 20.36GB

Qwen3.5 4B (Reasoning)阿里云

IQ 20.13GB

Gemma 4 26B A4B (Non-reasoning)GoogleMoE

IQ 20.115GB

Seed-OSS-36B-Instruct字节跳动

IQ 18.322GB

Qwen3 VL 32B (Reasoning)阿里云

IQ 17.920GB

HyperCLOVA X SEED Think (32B)Naver

IQ 17.019GB

Qwen3.5 4B (Non-reasoning)阿里云

IQ 16.03GB

Qwen3 30B A3B 2507 (Reasoning)阿里云MoE

IQ 15.818GB

GLM-4.7-Flash (Non-reasoning)Z AI (智谱 AI)MoE

IQ 15.519GB

Nemotron 3 Nano Omni 30B A3B ReasoningNVIDIAMoE

IQ 14.918GB

gpt-oss-20B (high)OpenAIMoE

IQ 14.913GB

gpt-oss-20B (low)OpenAIMoE

IQ 14.313GB

NVIDIA Nemotron 3 Nano 30B A3B (Reasoning)NVIDIAMoE

IQ 14.219GB

ZAYA1-8BZyphraMoE

IQ 14.15GB

Qwen3 Coder 30B A3B Instruct阿里云MoE

IQ 13.618GB

Tri-21B-think PreviewTrillion Labs

IQ 13.613GB

QwQ 32B阿里云

IQ 13.420GB

Qwen3 VL 30B A3B (Reasoning)阿里云MoE

IQ 13.318GB

Gemma 4 12B (Non-reasoning)Google

IQ 13.27GB

Devstral Small 2Mistral AI

IQ 13.114GB

Gemma 4 E4B (Reasoning)GoogleMoE

IQ 12.55GB

Tri-21B-ThinkTrillion Labs

IQ 12.413GB

Qwen3 4B 2507 (Reasoning)阿里云

IQ 12.02GB

MiniCPM5-1B (Reasoning)OpenBMB

IQ 12.01GB

Magistral Small 1.2Mistral AI

IQ 11.914GB

Devstral Small (May '25)Mistral AI

IQ 11.814GB

MiniCPM5-1B (Non-reasoning)OpenBMB

IQ 11.71GB

Qwen3 VL 32B Instruct阿里云

IQ 11.120GB

Ministral 3 14BMistral AI

IQ 11.18GB

DeepSeek R1 Distill Qwen 32BDeepSeek

IQ 11.019GB

Magistral Small 1Mistral AI

IQ 10.714GB

Qwen3 VL 8B (Reasoning)阿里云

IQ 10.65GB

EXAONE 4.0 32B (Reasoning)LG AI Research

IQ 10.619GB

Qwen3 32B (Reasoning)阿里云

IQ 10.520GB

DeepSeek R1 0528 Qwen3 8BDeepSeek

IQ 10.45GB

Qwen3.5 2B (Reasoning)阿里云

IQ 10.21GB

Qwen3 14B (Reasoning)阿里云

IQ 10.19GB

Nanbeige4.1-3B南北阁

IQ 10.12GB

Qwen3 VL 30B A3B Instruct阿里云MoE

IQ 10.018GB

DeepSeek R1 Distill Qwen 14BDeepSeek

IQ 9.88GB

Falcon-H1R-7BTII UAE

IQ 9.84GB

Qwen3 Omni 30B A3B (Reasoning)阿里云MoE

IQ 9.621GB

Step3 VL 10B阶跃星辰

IQ 9.56GB

Qwen3 30B A3B (Reasoning)阿里云MoE

IQ 9.318GB

Gemma 4 E2B (Reasoning)GoogleMoE

IQ 9.33GB

Devstral Small (Jul '25)Mistral AI

IQ 9.314GB

QwQ 32B-Preview阿里云

IQ 9.220GB

Qwen3 30B A3B 2507 Instruct阿里云MoE

IQ 9.118GB

Mistral Small 3.2Mistral AI

IQ 9.114GB

Ministral 3 8BMistral AI

IQ 9.05GB

NVIDIA Nemotron Nano 12B v2 VL (Reasoning)NVIDIA

IQ 9.08GB

Gemma 4 E4B (Non-reasoning)GoogleMoE

IQ 8.95GB

Granite 4.1 30BIBM

IQ 8.918GB

Qwen3.5 2B (Non-reasoning)阿里云

IQ 8.81GB

NVIDIA Nemotron 3 Nano 4BNVIDIA

IQ 8.82GB

NVIDIA Nemotron Nano 9B V2 (Reasoning)NVIDIA

IQ 8.85GB

Qwen3 32B (Non-reasoning)阿里云

IQ 8.620GB

Mistral Small 3.1Mistral AI

IQ 8.614GB

Llama 3.1 Nemotron Nano 4B v1.1 (Reasoning)NVIDIA

IQ 8.53GB

Qwen3 4B (Reasoning)阿里云

IQ 8.42GB

Qwen3 VL 8B Instruct阿里云

IQ 8.45GB

Olmo 3.1 32B ThinkAllen Institute for AI

IQ 8.119GB

Qwen3 VL 4B (Reasoning)阿里云

IQ 7.93GB

Qwen2.5 Instruct 32B阿里云

IQ 7.519GB

Qwen3 8B (Reasoning)阿里云

IQ 7.45GB

NVIDIA Nemotron 3 Nano 30B A3B (Non-reasoning)NVIDIAMoE

IQ 7.419GB

NVIDIA Nemotron Nano 9B V2 (Non-reasoning)NVIDIA

IQ 7.45GB

Qwen2.5 Coder Instruct 32B阿里云

IQ 7.119GB

Qwen3 4B 2507 Instruct阿里云

IQ 7.12GB

Qwen3 14B (Non-reasoning)阿里云

IQ 7.09GB

Mistral Small 3Mistral AI

IQ 6.914GB

Qwen3 30B A3B (Non-reasoning)阿里云MoE

IQ 6.818GB

Qwen3 4B (Non-reasoning)阿里云

IQ 6.82GB

Ministral 3 3BMistral AI

IQ 6.82GB

Granite 4.1 8BIBM

IQ 6.75GB

Sarvam 30B (high)SarvamMoE

IQ 6.619GB

Olmo 3.1 32B InstructAllen Institute for AI

IQ 6.519GB

Olmo 3 32B ThinkAllen Institute for AI

IQ 6.419GB

DeepSeek R1 Distill Llama 8BDeepSeek

IQ 6.45GB

Gemma 4 E2B (Non-reasoning)GoogleMoE

IQ 6.43GB

Solar MiniUpstage

IQ 6.26GB

Llama 3.1 Instruct 8BMeta

IQ 6.15GB

EXAONE 4.0 32B (Non-reasoning)LG AI Research

IQ 6.019GB

DeepHermes 3 - Mistral 24B Preview (Non-reasoning)Nous Research

IQ 5.314GB

Granite 4.0 H SmallIBMMoE

IQ 5.219GB

Qwen3 8B (Non-reasoning)阿里云

IQ 5.15GB

Qwen3 Omni 30B A3B Instruct阿里云MoE

IQ 5.121GB

OLMo 2 32BAllen Institute for AI

IQ 5.019GB

Qwen3.5 0.8B (Reasoning)阿里云

IQ 5.01GB

Phi-4Microsoft Azure

IQ 4.98GB

LFM2 24B A2BLiquid AIMoE

IQ 4.914GB

Gemma 3 27B InstructGoogle

IQ 4.816GB

Mistral Small (Sep '24)Mistral AI

IQ 4.713GB

Phi-3 Mini Instruct 3.8BMicrosoft Azure

IQ 4.62GB

Gemma 3n E4B Instruct Preview (May '25)GoogleMoE

IQ 4.65GB

NVIDIA Nemotron Nano 12B v2 VL (Non-reasoning)NVIDIA

IQ 4.68GB

Qwen2.5 Coder Instruct 7B 阿里云

IQ 4.55GB

Phi-4 Multimodal InstructMicrosoft Azure

IQ 4.53GB

Qwen3.5 0.8B (Non-reasoning)阿里云

IQ 4.41GB

Llama 2 Chat 7BMeta

IQ 4.34GB

Llama 3.2 Instruct 3BMeta

IQ 4.22GB

MiniCPM-V 4.6 1.3BOpenBMB

IQ 4.21GB

Jamba Reasoning 3BAI21 Labs

IQ 4.12GB

Qwen3 VL 4B Instruct阿里云

IQ 4.13GB

Reka Flash 3Reka AI

IQ 4.113GB

Olmo 3 7B ThinkAllen Institute for AI

IQ 4.04GB

OLMo 2 7BAllen Institute for AI

IQ 3.94GB

Molmo 7B-DAllen Institute for AI

IQ 3.85GB

Ling-mini-2.0蚂蚁 InclusionAIMoE

IQ 3.810GB

DeepSeek R1 Distill Qwen 1.5BDeepSeek

IQ 3.71GB

Gemma 3 12B InstructGoogle

IQ 3.47GB

Llama 3.2 Instruct 11B (Vision)Meta

IQ 3.37GB

Granite 4.1 3BIBM

IQ 3.22GB

DeepSeek Coder V2 Lite InstructDeepSeekMoE

IQ 3.110GB

Phi-4 Mini InstructMicrosoft Azure

IQ 3.02GB

DeepSeek LLM 67B Chat (V1)DeepSeek

IQ 3.04GB

Llama 2 Chat 13BMeta

IQ 3.08GB

OpenChat 3.5 (1210)OpenChat

IQ 3.04GB

Sarvam M (Reasoning)Sarvam

IQ 3.014GB

Exaone 4.0 1.2B (Reasoning)LG AI Research

IQ 2.91GB

Olmo 3 7B InstructAllen Institute for AI

IQ 2.84GB

Exaone 4.0 1.2B (Non-reasoning)LG AI Research

IQ 2.81GB

Granite 4.0 H 1BIBM

IQ 2.71GB

LFM2 2.6BLiquid AI

IQ 2.72GB

LFM2.5-1.2B-InstructLiquid AI

IQ 2.71GB

LFM2.5-1.2B-ThinkingLiquid AI

IQ 2.71GB

Qwen3 1.7B (Reasoning)阿里云

IQ 2.61GB

Granite 4.0 MicroIBM

IQ 2.42GB

DeepHermes 3 - Llama-3.1 8B Preview (Non-reasoning)Nous Research

IQ 2.35GB

Qwen Chat 14B阿里云

IQ 2.18GB

Command-R (Mar '24)Cohere

IQ 2.121GB

Granite 4.0 1BIBM

IQ 2.11GB

Mistral 7B InstructMistral AI

IQ 2.14GB

Molmo2-8BAllen Institute for AI

IQ 2.05GB

Granite 3.3 8B (Non-reasoning)IBM

IQ 1.85GB

LFM2 8B A1BLiquid AIMoE

IQ 1.85GB

Qwen3 1.7B (Non-reasoning)阿里云

IQ 1.51GB

Gemma 3n E4B InstructGoogleMoE

IQ 1.25GB

Llama 3 Instruct 8BMeta

IQ 1.25GB

Gemma 3 4B InstructGoogle

IQ 1.13GB

LFM2 1.2BLiquid AI

IQ 1.11GB

Llama 3.2 Instruct 1BMeta

IQ 1.11GB

Tiny Aya GlobalCohere

IQ 1.02GB

Gemma 3 1B InstructGoogle

IQ 1.01GB

Gemma 3n E2B InstructGoogleMoE

IQ 1.04GB

LFM2.5-VL-1.6BLiquid AI

IQ 1.01GB

Apertus 8B InstructSwiss AI Initiative

IQ 1.05GB

EXAONE 4.5 33B (Non-reasoning)LG AI Research

21GB

VRAM 需求为估算值（含 ~20% KV Cache 开销），实际值因推理框架（vLLM / llama.cpp / TGI）而异。 MoE 模型需全量加载权重，推理时仅激活部分参数。

开源模型选型指南

智能指数 vs VRAM 需求 (Q4)

GLM-5.2 (max)

MiniMax-M3

DeepSeek V4 Pro (Reasoning, Max Effort)

Kimi K2.6

MiMo-V2.5-Pro

Kimi K2.7 Code

Nex-N2-Pro

DeepSeek V4 Pro (Reasoning, High Effort)

DeepSeek V4 Flash (Reasoning, Max Effort)

GLM-5.1 (Reasoning)

MiMo-V2.5

GLM-5 (Reasoning)

MiniMax-M2.7

Kimi K2.5 (Reasoning)

Nemotron 3 Ultra 550B A55B (Reasoning)

DeepSeek V4 Flash (Reasoning, High Effort)

Qwen3.6 27B (Reasoning)

GLM-5.1 (Non-reasoning)

Kimi K2.6 (Non-reasoning)

Qwen3.5 27B (Reasoning)

GLM-4.7 (Reasoning)

Qwen3.5 397B A17B (Reasoning)

MiniMax-M2.5

Hy3-preview (Reasoning)

VRAM 兼容性检查器